• Robust masking approach isolating bare-soil pixels for SOM prediction. • LASSO-based image ranking improved feature selection and model efficiency. • Iterative modelling with ranked Sentinel-2 images enhanced prediction stability. • Integration of soil texture and spectral indices improved model accuracy. • Farm R²=0.83 shows strong local fit; Québec R²≈0.28–0.36 reflects realistic scaling. Predicting continuous soil properties from limited field observations remains a central challenge in precision agriculture, particularly when models must operate across multiple spatial scales. This study develops a multi-scale framework that combines multi-temporal Sentinel-2 imagery with legacy soil maps to estimate spatial variation in soil organic matter (SOM). Bare-soil pixels were extracted and spectral indices calculated after cloud and snow masking, and a LASSO-based procedure was used to select informative images before modelling. Two strategies were evaluated: one using only satellite-derived indices, and another integrating soil texture information. At the farm scale, twelve Sentinel-2 images yielded 422 field–image records. A simple field-level averaging baseline achieved RMSE = 0.14 log(%SOM) and R² = 0.74, while the hybrid model predicting at the zone level achieved RMSE = 0.16 log(%SOM) and R² = 0.83, capturing within-field variability despite slightly higher point-wise error. Remote-sensing-only models performed poorly (RMSE ≈ 0.35–0.37 log(%SOM) and R² < 0.10), demonstrating that spectral indices alone cannot represent subsurface conditions. The framework was then scaled to the province of Québec. Multi-year images, soil texture, topography, and climate variables were combined, and Random Forest, LightGBM, and CatBoost were tested after feature screening. At the province scale, predictive performance decreased (R² = 0.287–0.364), reflecting the increased agroclimatic, edaphic, and management heterogeneity across Québec. However, the corresponding RMSE values (≈ 0.101–0.103 log(%SOM)) indicate that prediction errors remain quantitatively moderate after back-transformation. Therefore, model performance at this scale should be interpreted not only in terms of explained variance, but also in terms of operational prediction error and regional differentiation capability. At the regional level, the framework supports decision-oriented applications that rely on relative field differentiation rather than precise point estimation.
Etezadi et al. (Sun,) studied this question.